Multidimensional BSS/WSS Criterion for Modified SBFS Feature Selection Method in Tumor Classification

Abstract

To avoid the defect of BSS/WSS criterion, we propose a multidimensional BSS/WSS feature selection criterion and modify the sequential backward floating selection (SBFS) algorithm to deal with the case where the covariance matrix is singular in this study. Then, we use support vector machine (SVM) to classify the gene expression data based on the proposed feature selection algorithm. The performance of the proposed approach is compared with BSS/WSS criterion and some other popular methods in feature selection and classification via the wellknown colon cancer and prostate datasets in microarray literature, which demonstrates that the proposed criterion can take into account genes' joint discriminatory power, and the proposed feature selection method can obtain correct and informative gene subset for tumor classification.